---
id: AGI-REQ-001
title: "Requirements for General Intelligence"
alternative_names:
  - "Capabilities Required for AGI"
  - "Ingredients of General Intelligence"
category: "Advanced Intelligence → General Intelligence"
volume: V
chapter: 28
entry_type: "Technical Article"
level: 4
status: "Draft"
last_reviewed: 2026-10-15
version: "0.1.0"
prerequisites:
  - "AGI-001"
  - "AGI-DEF-001"
related_entries:
  - "AGI-WM-001"
  - "AGI-CL-001"
  - "AGI-PATH-001"
  - "AGI-TEST-001"
  - "TECH-AGENT-001"
---

# Requirements for General Intelligence

## 1. Concise Definition
The **requirements for general intelligence** are the cognitive capabilities that researchers have proposed as necessary, or at least strongly facilitating, for a machine to count as generally intelligent: among them flexible generalisation, causal reasoning, world modelling, memory and continual learning, long-horizon planning and metacognition. Which of these are truly necessary, as opposed to merely useful, is unresolved, and nearly all claims of necessity in this literature are Contested or Hypothetical.

## 2. Overview & Significance
Definitions of AGI (see [AGI-DEF-001](../vol05_agi/AGI-DEF-001_definitions_and_operational_criteria_for_agi.md)) state a target. Requirements are the intermediate claims about what a system would have to be able to do to reach it. The distinction matters because the requirements drive research agendas and forecasts: if a capability such as persistent memory or causal modelling is necessary and current architectures lack it, then scaling alone would not suffice, and timelines that extrapolate from compute alone would be mistaken. If it is not necessary, the extrapolation may remain valid.

A requirement claim is of one of three kinds, and conflating them is a common source of confusion in the literature.

- **Necessity claims**: no system lacking capability $C$ can be generally intelligent. These are rarely proved, and for engineered systems usually cannot be, because a different mechanism might supply the same function.
- **Sufficiency claims**: any system with $C$ (at scale) will be generally intelligent. These are the claims made by scaling-based positions.
- **Diagnostic claims**: current systems fail at $C$, and the failure explains an observed limitation. These are empirical and can be tested.

This article lists the principal candidate requirements, states which kind of claim is made for each, and rates the evidence conservatively. It does not argue that any one list is correct.

## 3. Historical Development
Early AI treated general intelligence as a matter of reasoning over explicit representations (see [FOUND-003](../vol01_foundations/FOUND-003_knowledge_representation.md)), so the requirements were thought to be search, logic and knowledge. The connectionist and then deep-learning periods shifted attention to learning, and critiques from the cognitive-science tradition re-specified the requirements in terms of human cognition. Lake et al. (2017) argued that human-like learning relies on intuitive physics and psychology, compositional structure, and learning-to-learn, and that deep-network systems of the time lacked them. Marcus (2020) argued for hybrid systems with explicit knowledge, reasoning and cognitive models. Pearl (2009) provided the formal apparatus of causal inference that many requirement lists refer to.

The arrival of large pre-trained language models changed the debate without settling it. Some authors argue that such models partly exhibit capabilities previously listed as missing, whereas others argue that behavioural success does not show that the underlying capability is present; Mitchell and Krakauer (2023) review this disagreement. Bender and Koller (2020) argued that systems trained only on linguistic form cannot, in principle, learn meaning; this argument is itself contested. More recent assessment frameworks turn requirements into measurable domains: Hendrycks et al. (2025) list ten cognitive domains and report that current systems have critical deficits, in particular in long-term memory storage.

## 4. Technical / Conceptual Explanation

### 4.1 Candidate requirement 1: flexible generalisation and sample efficiency
Chollet (2019) characterises intelligence as the efficiency with which a system acquires new skills given priors and experience. A system that performs well only on tasks resembling its training distribution has skill but not necessarily generality. Lake et al. (2017) make a related point about sample efficiency: people learn new concepts from few examples, whereas the systems they reviewed required far more. Formally, one can express the gap through the *generalisation error under distribution shift*,
$$\mathcal{E}_{\text{shift}}(f) = \mathbb{E}_{(x,y)\sim Q}[\ell(f(x),y)] - \mathbb{E}_{(x,y)\sim P}[\ell(f(x),y)],$$
where $P$ is the training distribution and $Q$ the deployment distribution. Generality requires that this gap remain small across a wide set of $Q$. Whether any particular architecture achieves this is an empirical question.

Status: the claim that generalisation is *necessary* is part of nearly all definitions and is not seriously disputed in the sources reviewed; the claim that current systems *lack* it is Contested, because results depend on how novelty is measured (see [AGI-TEST-001](../vol05_agi/AGI-TEST-001_testing_claims_of_general_intelligence.md)).

### 4.2 Candidate requirement 2: causal and compositional reasoning
Pearl (2009) distinguishes three rungs of inference: association, intervention, and counterfactual reasoning. A system using only associations, $P(y\mid x)$, cannot in general answer interventional queries such as $P(y\mid \mathrm{do}(x))$ without additional assumptions. Lake et al. (2017) additionally emphasise compositionality, the recombination of known parts into new structures. Proponents argue that planning and scientific reasoning require both. Opponents argue that sufficiently large models acquire effective causal and compositional behaviour implicitly; the sources reviewed do not settle this. Status: Contested.

### 4.3 Candidate requirement 3: world models
A world model is an internal predictive model of how an environment changes in response to actions. LeCun (2022) proposes an architecture in which a learned world model is central to planning; reinforcement-learning work such as Ha and Schmidhuber (2018) and Hafner et al. (2023) uses learned models for planning in simulated environments. Whether general intelligence *requires* an explicit world model is Contested, as is whether language models contain one. The topic has its own article: [AGI-WM-001](../vol05_agi/AGI-WM-001_world_models.md).

### 4.4 Candidate requirement 4: memory and continual learning
A generally intelligent agent presumably accumulates knowledge over time. Standard neural-network training, however, suffers from catastrophic forgetting when trained sequentially (Kirkpatrick et al., 2017), and Dohare et al. (2024) report that standard deep-learning methods can lose the ability to keep learning in long continual-learning settings. Hendrycks et al. (2025) identify long-term memory storage as a critical deficit of current systems. Deployed workarounds (large context windows, external memory) differ from updating the model's own parameters. Status: the diagnostic claim that current systems lack durable, updatable memory is Well Supported in the narrow sense that model weights are fixed after training in standard deployments; whether this blocks AGI is Contested. See [AGI-CL-001](../vol05_agi/AGI-CL-001_continual_learning_and_persistent_memory.md).

### 4.5 Candidate requirement 5: long-horizon planning and agency
Many economically valuable tasks require sustained, self-directed action. Kwa et al. (2025) propose measuring AI progress by the length of tasks, in human-expert time, that models complete with 50% success; they report that for frontier models such as Claude 3.7 Sonnet this horizon was around 50 minutes at the time of the paper, and that the horizon has been doubling approximately every seven months since 2019 in their data. These figures are specific to the task suite, success threshold and date of the paper; they are Preliminary and are not a statement that long-horizon agency has been achieved or will be. Agent architectures are covered in [TECH-AGENT-001](../vol03_technical/TECH-AGENT-001_autonomous_agents.md).

### 4.6 Candidate requirement 6: metacognition and calibration
Morris et al. (2024) include metacognitive tasks within the scope of the cognitive abilities relevant to AGI: a system should be able to monitor and report on its own knowledge and uncertainty. Calibration (agreement between stated confidence and accuracy) is measurable, but whether it is *required* for generality is not established. Status: Hypothetical as a requirement; calibration as a measurable property is not in dispute.

### 4.7 Candidate requirement 7: embodiment and physical grounding
One position holds that general intelligence requires sensorimotor interaction with the physical world, so that concepts are grounded. Another holds that cognitive generality is separable from physical competence, and Morris et al. (2024) deliberately focus on cognitive and metacognitive rather than physical tasks. Bender and Koller (2020) argue from the form/meaning distinction that grounding is needed for meaning. The question of whether grounding requires a body or only some form of interaction with an environment is open. Status: Contested.

## 5. Types, Components & Classifications
Requirement proposals can be sorted along two axes: whether they are *capabilities* (what the system can do) or *mechanisms* (how it does it), and whether they are claimed as *necessary* or *sufficient*. The Morris et al. (2024) principle of focusing on capabilities rather than processes implies that mechanism-level requirements, such as "must contain an explicit world model", are not part of the definition of AGI; they are hypotheses about how to achieve it.

| Candidate | Type | Kind of claim | Main sources | Status |
| :--- | :--- | :--- | :--- | :--- |
| Flexible generalisation, sample efficiency | Capability | Necessity (broadly accepted); diagnostic (contested) | Chollet (2019); Lake et al. (2017) | Contested (as a diagnosis of current systems) |
| Causal and compositional reasoning | Capability and mechanism | Necessity (disputed) | Pearl (2009); Marcus (2020) | Contested |
| World model | Mechanism | Necessity (disputed) | LeCun (2022); Ha & Schmidhuber (2018) | Contested |
| Persistent memory, continual learning | Capability | Diagnostic; necessity disputed | Kirkpatrick et al. (2017); Dohare et al. (2024) | Contested |
| Long-horizon agency | Capability | Necessity for economic definitions | Kwa et al. (2025) | Preliminary |
| Metacognition, calibration | Capability | Hypothetical | Morris et al. (2024) | Hypothetical |
| Embodiment | Mechanism | Necessity (disputed) | Bender & Koller (2020) | Contested |

## 6. Applications & Concrete Examples
Requirement lists are applied in three ways.

- **Gap analyses.** Hendrycks et al. (2025) score named models across ten domains and conclude that deficits remain, particularly in long-term memory storage. The scores are the authors' own and Preliminary.
- **Benchmark design.** A requirement can be turned into a test. ARC-AGI-2 (Chollet et al., 2025) targets generalisation to novel tasks; the METR time-horizon work (Kwa et al., 2025) targets long-horizon agency in software tasks. Neither covers all candidate requirements.
- **Research programmes.** The pathways in [AGI-PATH-001](../vol05_agi/AGI-PATH-001_pathways_to_agi.md) can be read as bets on different requirements: scaling bets on sufficiency of a single mechanism, world-model approaches on requirement 3, neuro-symbolic approaches on requirement 2.

No deployed system has been demonstrated to satisfy a full list of these requirements. Demonstrations are partial, and laboratory demonstrations should not be read as deployed capability.

## 7. Advantages, Limitations & Unresolved Questions
**Value of requirement lists.** They turn a vague target into testable sub-claims and identify which technical gaps might matter most.

**Limitations.**
- A list inferred from human cognition may include features that are artefacts of biology rather than needs of general intelligence.
- Capabilities can be behaviourally present without the proposed mechanism, so a failure of a mechanism-level requirement does not imply failure of the capability.
- Lists are rarely derived from a theory that says why these items and not others; they are largely assembled from cognitive science and from observed AI failures. That makes completeness impossible to establish.
- Diagnostic claims about current systems age quickly, because systems change. Any specific failure documented in a paper reflects the models tested at that date.

**Open questions.**
1. Is any single capability (for example, continual learning) a bottleneck for the rest?
2. Can capabilities be tested independently, or does interaction between them make the sum more than the parts?
3. Do requirements differ by definition? For an economic definition, long-horizon agency and reliability weigh heavily; for a skill-acquisition definition, sample efficiency does.
4. Whether understanding, in the sense contested by Bender and Koller (2020) and by Mitchell and Krakauer (2023), is a separate requirement or an emergent description of the others.

## 8. Current State of Knowledge
As of 2026-10-15 there is no agreed list of necessary capabilities for general intelligence, and no experiment has shown that any one candidate is strictly necessary. The sources reviewed support three limited statements. First, standard training pipelines in deployed language models produce fixed weights, and sequential training of neural networks has a documented forgetting problem (Kirkpatrick et al., 2017). Second, benchmark and framework papers identify gaps in current systems, including long-term memory (Hendrycks et al., 2025) and long-horizon task completion (Kwa et al., 2025), but report them on specific suites and dates. Third, whether the gaps are fundamental or will be closed by existing methods is Contested; Mitchell and Krakauer (2023) summarise opposing views on whether such systems "understand". Positions on this question are not distributed evenly and are not a consensus. For current system assessments, consult the Living Reference collections (`MODEL-001`, `FACT-001`) rather than this article.

## 9. Future Trajectory & Strategic Outlook
This section is descriptive and registers no forecast.

- **Near term (0 to 3 years).** Expect further work that turns individual requirements into measurable sub-benchmarks and gap analyses (for memory, planning, and generalisation), and continued disagreement over which are decisive.
- **Medium term (3 to 10 years).** Empirical results will bear on diagnostic claims: if systems close particular gaps without architectural change, the corresponding necessity claims weaken; if gaps persist as systems scale, the claims strengthen. The sources reviewed contain no evidence on which outcome will occur.
- **Long term (10+ years).** Requirement lists may be replaced by capability profiles measured directly, as the Morris et al. (2024) emphasis on capabilities over processes suggests. This is speculative.

Dated, testable predictions about these questions belong in the Forecast Register (`FCST-REG-001`).

## 10. Related Entries & Knowledge Graph
- **Foundational Concepts**:
  - [[AGI-001] Artificial General Intelligence](../vol05_agi/AGI-001_artificial_general_intelligence.md): Overview of AGI, including a summary list of cognitive requirements.
  - [[AGI-DEF-001] Definitions and Operational Criteria for AGI](../vol05_agi/AGI-DEF-001_definitions_and_operational_criteria_for_agi.md): The targets that requirements are meant to serve.
  - [[FOUND-003] Knowledge Representation](../vol01_foundations/FOUND-003_knowledge_representation.md): Symbolic and structured representations.
- **Related Methods & Technologies**:
  - [[AGI-WM-001] World Models](../vol05_agi/AGI-WM-001_world_models.md): Requirement 3 in detail.
  - [[AGI-CL-001] Continual Learning and Persistent Memory](../vol05_agi/AGI-CL-001_continual_learning_and_persistent_memory.md): Requirement 4 in detail.
  - [[TECH-AGENT-001] Autonomous AI Agents](../vol03_technical/TECH-AGENT-001_autonomous_agents.md): Agent architectures for long-horizon action.
  - [[TECH-RL-001] Reinforcement Learning](../vol03_technical/TECH-RL-001_reinforcement_learning.md): Learning from interaction.
- **Advanced Implications & Future Trajectories**:
  - [[AGI-PATH-001] Pathways to AGI](../vol05_agi/AGI-PATH-001_pathways_to_agi.md): Research programmes built on different requirement hypotheses.
  - [[AGI-TEST-001] Testing Claims of General Intelligence](../vol05_agi/AGI-TEST-001_testing_claims_of_general_intelligence.md): How requirements are turned into tests.
  - [[CAP-REASON-001] Reasoning, Planning and Long-Horizon Capability](../vol04_capabilities/CAP-REASON-001_reasoning_planning_and_long_horizon_capability.md): Capability evidence for requirements 2 and 5.

## 11. References & Foundational Literature
1. Bender, E. M., & Koller, A. (2020). *Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data*. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5185–5198. https://doi.org/10.18653/v1/2020.acl-main.463. [Argues that systems trained only on linguistic form cannot learn meaning; a contested position.]
2. Chollet, F. (2019). *On the Measure of Intelligence*. arXiv preprint. arXiv:1911.01547. [Defines intelligence as skill-acquisition efficiency; basis for requirement 1.]
3. Chollet, F., Knoop, M., Kamradt, G., Landers, B., & Pinkard, H. (2025). *ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems*. arXiv preprint. arXiv:2505.11831. [Benchmark designed around generalisation to novel tasks.]
4. Dohare, S., Hernandez-Garcia, J. F., Lan, Q., Rahman, P., Mahmood, A. R., & Sutton, R. S. (2024). *Loss of plasticity in deep continual learning*. Nature, 632(8026), 768–774. https://doi.org/10.1038/s41586-024-07711-7. [Reports loss of ability to keep learning in long continual-learning settings with standard methods.]
5. Ha, D., & Schmidhuber, J. (2018). *World Models*. arXiv preprint. arXiv:1803.10122. [Learned generative world model used for control in simulated environments.]
6. Hafner, D., Pasukonis, J., Ba, J., & Lillicrap, T. (2023). *Mastering Diverse Domains through World Models*. arXiv preprint. arXiv:2301.04104. [Model-based reinforcement learning across many tasks with one configuration, as reported by the authors.]
7. Hendrycks, D., Song, D., Szegedy, C., Lee, H., Gal, Y., Brynjolfsson, E., et al. (2025). *A Definition of AGI*. arXiv preprint. arXiv:2510.18212. [Ten-domain framework; abstract states current systems have deficits in long-term memory storage.]
8. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., et al. (2017). *Overcoming catastrophic forgetting in neural networks*. Proceedings of the National Academy of Sciences, 114(13), 3521–3526. https://doi.org/10.1073/pnas.1611835114. [Documents catastrophic forgetting and proposes elastic weight consolidation.]
9. Kwa, T., West, B., Becker, J., Deng, A., Garcia, K., Hasin, M., et al. (2025). *Measuring AI Ability to Complete Long Software Tasks*. NeurIPS 2025; arXiv:2503.14499. [50% time-horizon metric and reported doubling time, quoted with their scope.]
10. Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017). *Building machines that learn and think like people*. Behavioral and Brain Sciences, 40. https://doi.org/10.1017/S0140525X16001837. [Proposes intuitive physics and psychology, compositionality and learning-to-learn as ingredients missing from deep-network systems of that time.]
11. LeCun, Y. (2022). *A Path Towards Autonomous Machine Intelligence* (version 0.9.2, 2022-06-27). OpenReview position paper. https://openreview.net/forum?id=BZ5a1r-kVsf. [Proposes a modular architecture built around a learned world model. [VERIFICATION NEEDED: OpenReview page not directly retrieved; version and date confirmed only via secondary listings]]
12. Marcus, G. (2020). *The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence*. arXiv preprint. arXiv:2002.06177. [Argues for hybrid knowledge-driven and learning-based systems.]
13. Mitchell, M., & Krakauer, D. C. (2023). *The debate over understanding in AI's large language models*. Proceedings of the National Academy of Sciences, 120(13). https://doi.org/10.1073/pnas.2215907120. [Surveys opposing positions on whether language models understand.]
14. Morris, M. R., Sohl-Dickstein, J., Fiedel, N., Warkentin, T., Dafoe, A., Faust, A., Farabet, C., & Legg, S. (2024). *Levels of AGI for Operationalizing Progress on the Path to AGI*. Proceedings of ICML 2024; arXiv:2311.02462. [Capability-focused principles, including cognitive and metacognitive scope.]
15. Pearl, J. (2009). *Causality: Models, Reasoning, and Inference* (2nd ed.). Cambridge University Press. ISBN: 978-0521895606. [Formal treatment of interventions and counterfactuals.]

## 12. Evidential Classification Matrix
| Section / Claim | Evidence Category | Supporting Source | Confidence / Status |
| :--- | :--- | :--- | :--- |
| Observational association does not in general determine interventional effects | Established (mathematical result) | Pearl (2009) | High |
| Sequential neural-network training can cause catastrophic forgetting | Established (replicated empirical phenomenon) | Kirkpatrick et al. (2017) | High |
| Standard deep learning can lose plasticity in long continual settings | Preliminary (single study, Nature-published; replication ongoing) | Dohare et al. (2024) | Moderate |
| Current systems have a critical deficit in long-term memory storage | Preliminary (authors' own scoring) | Hendrycks et al. (2025) | Moderate |
| 50% time horizon about 50 minutes; doubling about every seven months | Preliminary (task-suite specific) | Kwa et al. (2025) | Moderate |
| Causal reasoning, world models, memory or embodiment are necessary for AGI | Contested | Marcus (2020); LeCun (2022); Bender & Koller (2020); Mitchell & Krakauer (2023) | Active debate |
| Metacognition is required for AGI | Hypothetical | Morris et al. (2024) | Low |

## 13. Revision History
- **v0.1.0** (2026-10-15): Initial draft. Citations checked against arXiv/Crossref/publisher pages where marked; not independently fact-checked.